feat: add LLM retry and fallback

Made-with: Cursor

src/config/env.ts

@@ -14,9 +14,11 @@ const envSchema = z.object({
 
   LLM_BASE_URL: z.string().url().default('http://localhost:11434/v1'),
   LLM_MODEL: z.string().default('qwen2.5:14b'),
+  LLM_FALLBACK_MODEL: z.string().optional(),
   LLM_API_KEY: z.string().optional(),
   LLM_TIMEOUT_MS: z.coerce.number().default(15000),
   LLM_MAX_RETRIES: z.coerce.number().min(0).default(1),
+  LLM_RETRY_DELAY_MS: z.coerce.number().min(0).default(1000),
   LLM_TEMPERATURE: z.coerce.number().min(0).max(2).default(0.7),
   LLM_MAX_TOKENS: z.coerce.number().default(2048),
 

src/services/llm/llm.service.ts

@@ -5,10 +5,13 @@ import type { Stack, Level, QuestionType } from '../../db/schema/enums.js';
 export interface LlmConfig {
   baseUrl: string;
   model: string;
+  fallbackModel?: string;
   apiKey?: string;
   timeoutMs: number;
   temperature: number;
   maxTokens: number;
+  maxRetries: number;
+  retryDelayMs: number;
 }
 
 export interface ChatMessage {
@@ -61,14 +64,42 @@ export class LlmService {
     this.config = {
       baseUrl: config?.baseUrl ?? env.LLM_BASE_URL,
       model: config?.model ?? env.LLM_MODEL,
+      fallbackModel: config?.fallbackModel ?? env.LLM_FALLBACK_MODEL,
       apiKey: config?.apiKey ?? env.LLM_API_KEY,
       timeoutMs: config?.timeoutMs ?? env.LLM_TIMEOUT_MS,
       temperature: config?.temperature ?? env.LLM_TEMPERATURE,
       maxTokens: config?.maxTokens ?? env.LLM_MAX_TOKENS,
+      maxRetries: config?.maxRetries ?? env.LLM_MAX_RETRIES,
+      retryDelayMs: config?.retryDelayMs ?? env.LLM_RETRY_DELAY_MS,
     };
   }
 
   async chat(messages: ChatMessage[]): Promise<string> {
+    let lastError: Error | null = null;
+
+    const modelsToTry = [this.config.model];
+    if (this.config.fallbackModel) {
+      modelsToTry.push(this.config.fallbackModel);
+    }
+
+    for (const model of modelsToTry) {
+      for (let attempt = 0; attempt <= this.config.maxRetries; attempt++) {
+        try {
+          return await this.executeChat(messages, model);
+        } catch (err) {
+          lastError = err instanceof Error ? err : new Error('LLM request failed');
+          if (attempt < this.config.maxRetries) {
+            const delayMs = this.config.retryDelayMs * Math.pow(2, attempt);
+            await sleep(delayMs);
+          }
+        }
+      }
+    }
+
+    throw lastError ?? new Error('LLM request failed');
+  }
+
+  private async executeChat(messages: ChatMessage[], model: string): Promise<string> {
     const url = `${this.config.baseUrl.replace(/\/$/, '')}/chat/completions`;
 
     const headers: Record<string, string> = {
@@ -79,7 +110,7 @@ export class LlmService {
     }
 
     const body = {
-      model: this.config.model,
+      model,
       messages: messages.map((m) => ({ role: m.role, content: m.content })),
       temperature: this.config.temperature,
       max_tokens: this.config.maxTokens,
@@ -161,6 +192,10 @@ Rules: type must be one of: ${typeList}. For single_choice/multiple_select: opti
   }
 }
 
+function sleep(ms: number): Promise<void> {
+  return new Promise((resolve) => setTimeout(resolve, ms));
+}
+
 function extractJson(text: string): string {
   const trimmed = text.trim();
   const match = trimmed.match(/\{[\s\S]*\}/);